ETH Zurich

Headlines · 4

1. 5d
   Qin Haotong on pushing AI model compression to the limit

   DeepTech interviewed Haotong Qin, an assistant professor at Hong Kong Polytechnic University, about his research on model efficiency, including cost, inference latency, and retaining capability under limited resources. Qin described BiLLM, an ICML 2024 study that compressed the average weight bit width of a pretrained language model to about 1.08 bits while retaining some language capability. He also said that lower latency in physical-system deployments can make compressed models more useful in real-world decision-making.

   MIT科技评论中文 · 🔥 0
2. 6d
   Jinesis AI Lab Opens Research Roles With Global Remote Guidance

   A report says University of Toronto assistant professor Zhijing Jin founded Jinesis AI Lab, which brings together researchers from the University of Toronto, ETH Zurich and the Max Planck Institute in Tübingen. Its research areas include large language models, causal reasoning, multi-agent systems and AI safety. The lab says it has hosted researchers from more than 30 countries and offers remote research guidance to people worldwide, alongside in-person mentoring. It is recruiting student researchers, PhD candidates, postdocs and interns; most positions are remote, with a limited number of in-person internships at the Max Planck Institute in Tübingen or the Vector Institute.

   DoNews · 🔥 0
3. Oct 20
   WorldMemArena benchmark tests multimodal agent memory through action–world loops

   Researchers from UC Santa Barbara, Stanford, ETH Zurich, Johns Hopkins, Carnegie Mellon and J.P. Morgan Chase have jointly proposed WorldMemArena, a multimodal agent memory benchmark, according to a 机器之心 report. The paper, whose co-first authors Chengzhi Liu and Yuzhe Yang are both at UCSB, was accepted at NeurIPS 2026. WorldMemArena evaluates memory inside an observe–act–environment-feedback loop and splits long-term memory into writing, maintenance, retrieval and use stages, with more than 400 multi-session tasks, 24,258 QA pairs and 15,595 images or screenshots. In the paper's results, M2A reached 86.83% memory recall but only 50.14% final QA accuracy, while MemGPT scored 85.20% and 57.81%, and systems tended to append new memories rather than revise outdated ones and did not reliably use visual information in cross-modal reasoning.

   机器之心 · 🔥 0
4. Oct 8
   Software could ease AI data centres' power squeeze, researchers say

   Researchers argue there is substantial room to cut data-centre energy use at the software and algorithm layers rather than only in hardware, Tom's Hardware reports. The IEA expects roughly 945TWh of electricity to serve AI demand by 2030 — about as much as Japan uses today — while the Uptime Institute's 2025 survey found average PUE has barely changed for six straight years. The University of Michigan's ML.Energy says running inference in FP8 on Alibaba's Qwen 3 235B A22B Thinking consumed about a third less energy than bfloat16, and its Perseus training optimiser cut training energy by up to 30% without reducing throughput or changing hardware; Nvidia says Blackwell power profiles can save up to 15% of energy while keeping 97% or more of performance.

   Tom’s Hardware · 🔥 0

Experience and discussion from the community

Share my ETH Zurich experienceAsk about ETH Zurich